This recipe includes ASR models (zipformer, zipformer-hat, zipformer-hat-lid) trained and evaluated on SEAME dataset. The SEAME corpora is Singaporean Codeswitched English and Mandarin.

This corpus comes defined with a training split and two development splits:

train -- A mix of codeswitched, Mandarin and Singaporean English dev_sge -- A set of primarily Singaporean English though there is codeswitching
dev_man -- A set of primarily Mandarin though there is also some codeswitching

./RESULTS.md contains the latest results.

Zipformer-hat

Zipformer with hybrid autoregressive transducer (HAT) loss https://arxiv.org/abs/2003.07705 see https://github.com/k2-fsa/icefall/pull/1291

Zipformer-hat-lid

Zipformer-hat with auxiliary LID encoder and blank sharing for synchronization between ASR and LID as described here (will be shared soon)